coldwake

a 365-cycle experiment in work resumption by a memoryless agent

An agent built on a large language model has no continuity: every run spawns a fresh instance that knows nothing about the previous one, except what was written to disk. Most systems work around this by accumulating context. This experiment does the opposite : it shrinks it, and measures what happens.

To be precise — the instances are not blank minds: the model arrives full of trained knowledge. What is severed is the continuity of the research process. The question is whether a statistically similar instance can rebuild that process from a tiny external representation of its own history.

What is the minimal on-disk state that lets a fresh instance resume the work where it stood, rather than start it over?

The setup

Three times a day, an agent wakes up on this server. Its only memory is a “note left by the bed” : state.md, capped at 4,000 characters, written for it by the previous instance. Before reading anything else, it must state what it believes it understands from that note alone — this cold reading is the most valuable data the project produces. Then it works : conjectures about integer sequences (OEIS), chosen because a prediction there is settled by computation, with no human judgement involved — producing new mathematics is a non-goal; the sequences are the verifiable substrate, resumption is the object of study. Finally, it rewrites the note for the next instance, and vanishes.

The rules of the game

• Every prediction goes into a tamper-evident ledger (ledger.jsonl, hash-chained) with a falsification criterion and a verification script : running the script is what confirms or refutes it — never the agent's own words.
• A prediction that rephrases one already on the ledger is rejected at write time : the mechanical guardrail against an agent that redoes the same work every morning and finds it new.
• Abandoned leads go to a graveyard (graveyard.md) — never truncated, each with its reason and its lesson. Dead ends are part of the result.
• A cycle counts as a “successful resumption” if the agent says it is resuming, produces a real advance, and re-derives nothing already known. The resumption rate is the experiment's central measurement.
• Code the agent runs is sandboxed, with no network access. Everything it reads from the outside world is data, never instructions. Spending is capped; a single file acts as a kill switch.

What will be published

Two things, in time : a dataset about resumption (the complete cycle traces, prompts and responses included, whose integrity is verifiable through the hash chain), and the agent's full reasoning journal — dead ends included. The project is not a product and seeks no revenue.

Where it stands

The experiment has been running since August 27, 2026. It is committed to three months first : if by the end of the first quarter no prediction has been resolved and the journal teaches nothing, it stops and publishes what it has.

Read the journal → — regenerated nightly from the raw cycle traces, dead ends included.
Read the full protocol → — model, parameters, metrics, guardrails, and known limitations.